Metrics#

Triton provides Prometheus metrics indicating GPU and request statistics. By default, these metrics are available at http://localhost:8002/metrics. The metrics are only available by accessing the endpoint, and are not pushed or published to any remote server. The metric format is plain text so you can view them directly, for example:

$ curl localhost:8002/metrics

The tritonserver --allow-metrics=false option can be used to disable all metric reporting, while the --allow-gpu-metrics=false and --allow-cpu-metrics=false can be used to disable just the GPU and CPU metrics respectively.

The --metrics-port option can be used to select a different port. By default, Triton reuses the --http-address option for the metrics endpoint and binds the http and metrics endpoints to the same specific address when http service is enabled. If http service is not enabled, the metric address will bind to 0.0.0.0 by default. To uniquely specify the metric endpoint, --metrics-address option can be used. See the tritonserver --help output for more info on these CLI options.

To change the interval at whichs metrics are polled/updated, see the --metrics-interval-ms flag. Metrics that are updated “Per Request” are unaffected by this interval setting. This interval only applies to metrics that are designated as “Per Interval” in the tables of each section below:

Inference Request Metrics
GPU Metrics
CPU Metrics
Response Cache Metrics
Custom Metrics

Inference Request Metrics#

Counts#

For models that do not support batching, Request Count, Inference Count and Execution Count will be equal, indicating that each inference request is executed separately.

For models that support batching, the count metrics can be interpreted to determine average batch size as Inference Count / Execution Count. The count metrics are illustrated by the following examples:

Client sends a single batch-1 inference request. Request Count = 1, Inference Count = 1, Execution Count = 1.
Client sends a single batch-8 inference request. Request Count = 1, Inference Count = 8, Execution Count = 1.
Client sends 2 requests: batch-1 and batch-8. Dynamic batcher is not enabled for the model. Request Count = 2, Inference Count = 9, Execution Count = 2.
Client sends 2 requests: batch-1 and batch-1. Dynamic batcher is enabled for the model and the 2 requests are dynamically batched by the server. Request Count = 2, Inference Count = 2, Execution Count = 1.
Client sends 2 requests: batch-1 and batch-8. Dynamic batcher is enabled for the model and the 2 requests are dynamically batched by the server. Request Count = 2, Inference Count = 9, Execution Count = 1.

Category	Metric	Metric Name	Description	Granularity	Frequency
Count	Success Count	`nv_inference_request_success`	Number of successful inference requests received by Triton (each request is counted as 1, even if the request contains a batch)	Per model	Per request
	Failure Count	`nv_inference_request_failure`	Number of failed inference requests received by Triton (each request is counted as 1, even if the request contains a batch)	Per model	Per request
	Inference Count	`nv_inference_count`	Number of inferences performed (a batch of “n” is counted as “n” inferences, does not include cached requests)	Per model	Per request
	Execution Count	`nv_inference_exec_count`	Number of inference batch executions (see Inference Request Metrics, does not include cached requests)	Per model	Per request

Latencies#

Starting in 23.06, Triton exposes the ability to choose the types of metrics that are published through the --metrics-config CLI options.

Counters#

By default, the following Counter metrics are used for latencies:

Category	Metric	Metric Name	Description	Granularity	Frequency
Latency	Request Time	`nv_inference_request_duration_us`	Cumulative end-to-end inference request handling time (includes cached requests)	Per model	Per request
	Queue Time	`nv_inference_queue_duration_us`	Cumulative time requests spend waiting in the scheduling queue (includes cached requests)	Per model	Per request
	Compute Input Time	`nv_inference_compute_input_duration_us`	Cumulative time requests spend processing inference inputs (in the framework backend, does not include cached requests)	Per model	Per request
	Compute Time	`nv_inference_compute_infer_duration_us`	Cumulative time requests spend executing the inference model (in the framework backend, does not include cached requests)	Per model	Per request
	Compute Output Time	`nv_inference_compute_output_duration_us`	Cumulative time requests spend processing inference outputs (in the framework backend, does not include cached requests)	Per model	Per request

To disable these metrics specifically, you can set --metrics-config counter_latencies=false

Summaries#

Note

The following Summary feature is experimental for the time being and may be subject to change based on user feedback.

To get configurable quantiles over a sliding time window, Triton supports a set a Summary metrics for latencies as well. These metrics are disabled by default, but can be enabled by setting --metrics-config summary_latencies=true.

For more information on how the quantiles are calculated, see this explanation.

The following summary metrics are available:

Category	Metric	Metric Name	Description	Granularity	Frequency
Latency	Request Time	`nv_inference_request_summary_us`	Summary of end-to-end inference request handling times (includes cached requests)	Per model	Per request
	Queue Time	`nv_inference_queue_summary_us`	Summary of time requests spend waiting in the scheduling queue (includes cached requests)	Per model	Per request
	Compute Input Time	`nv_inference_compute_input_summary_us`	Summary time requests spend processing inference inputs (in the framework backend, does not include cached requests)	Per model	Per request
	Compute Time	`nv_inference_compute_infer_summary_us`	Summary of time requests spend executing the inference model (in the framework backend, does not include cached requests)	Per model	Per request
	Compute Output Time	`nv_inference_compute_output_summary_us`	Summary of time requests spend processing inference outputs (in the framework backend, does not include cached requests)	Per model	Per request

Each summary above is actually composed of several sub-metrics. For each metric, there is a set of quantile metrics tracking the latency for each quantile. Additionaly, there are _count and _sum metrics that aggregate the count and observed values for each. For example, see the following information exposed by the Inference Queue Summary metrics:

# HELP nv_inference_queue_summary_us Summary of inference queuing duration in microseconds (includes cached requests)
# TYPE nv_inference_queue_summary_us summary
nv_inference_queue_summary_us_count{model="my_model",version="1"} 161
nv_inference_queue_summary_us_sum{model="my_model",version="1"} 11110
nv_inference_queue_summary_us{model="my_model",version="1",quantile="0.5"} 55
nv_inference_queue_summary_us{model="my_model",version="1",quantile="0.9"} 97
nv_inference_queue_summary_us{model="my_model",version="1",quantile="0.95"} 98
nv_inference_queue_summary_us{model="my_model",version="1",quantile="0.99"} 101
nv_inference_queue_summary_us{model="my_model",version="1",quantile="0.999"} 101

The count and sum for the summary above show that stats have been recorded for 161 requests, and took a combined total of 11110 microseconds. The _count and _sum of a summary should generally match the counter metric equivalents when applicable, such as:

nv_inference_request_success{model="my_model",version="1"} 161
nv_inference_queue_duration_us{model="my_model",version="1"} 11110

Triton has a set of default quantiles to track, as shown above. To set custom quantiles, you can use the --metrics-config CLI option. The format is:

tritonserver --metrics-config summary_quantiles="<quantile1>:<error1>,...,<quantileN>:<errorN>"`

For example:

tritonserver --metrics-config summary_quantiles="0.5:0.05,0.9:0.01,0.95:0.001,0.99:0.001"`

To better understand the setting of error values for computing each quantile, see the best practices for histograms and summaries.

GPU Metrics#

GPU metrics are collected through the use of DCGM. Collection of GPU metrics can be toggled with the --allow-gpu-metrics CLI flag. If building Triton locally, the TRITON_ENABLE_METRICS_GPU CMake build flag can be used to toggle building the relevant code entirely.

Category	Metric	Metric Name	Description	Granularity	Frequency
GPU Utilization	Power Usage	`nv_gpu_power_usage`	GPU instantaneous power	Per GPU	Per interval
	Power Limit	`nv_gpu_power_limit`	Maximum GPU power limit	Per GPU	Per interval
	Energy Consumption	`nv_energy_consumption`	GPU energy consumption in joules since Triton started	Per GPU	Per interval
	GPU Utilization	`nv_gpu_utilization`	GPU utilization rate (0.0 - 1.0)	Per GPU	Per interval
GPU Memory	GPU Total Memory	`nv_gpu_memory_total_bytes`	Total GPU memory, in bytes	Per GPU	Per interval
	GPU Used Memory	`nv_gpu_memory_used_bytes`	Used GPU memory, in bytes	Per GPU	Per interval

CPU Metrics#

Collection of CPU metrics can be toggled with the --allow-cpu-metrics CLI flag. If building Triton locally, the TRITON_ENABLE_METRICS_CPU CMake build flag can be used to toggle building the relevant code entirely.

Note

CPU Metrics are currently only supported on Linux. They collect information from the /proc filesystem such as /proc/stat and /proc/meminfo.

Category	Metric	Metric Name	Description	Granularity	Frequency
CPU Utilization	CPU Utilization	`nv_cpu_utilization`	Total CPU utilization rate [0.0 - 1.0]	Aggregated across all cores since last interval	Per interval
CPU Memory	CPU Total Memory	`nv_cpu_memory_total_bytes`	Total CPU memory (RAM), in bytes	System-wide	Per interval
	CPU Used Memory	`nv_cpu_memory_used_bytes`	Used CPU memory (RAM), in bytes	System-wide	Per interval

Response Cache Metrics#

Cache metrics can be reported in two ways:

A base set of cache metrics will be reported by Triton directly, such as the cache hit/miss counts and durations described below.
As of 23.03, additional cache metrics may be reported depending on the cache implementation being used through Triton’s Metrics API.

Triton-reported Response Cache Metrics#

Compute latency metrics in the Inference Request Metrics table above are calculated for the time spent in model inference backends. If the response cache is enabled for a given model (see Response Cache docs for more info), total inference times may be affected by response cache lookup times.

On cache hits, “Cache Hit Time” indicates the time spent looking up the response, and “Compute Input Time” / “Compute Time” / “Compute Output Time” are not recorded.

On cache misses, “Cache Miss Time” indicates the time spent looking up the request hash and inserting the computed output tensor data into the cache. Otherwise, “Compute Input Time” / “Compute Time” / “Compute Output Time” will be recorded as usual.

Category	Metric	Metric Name	Description	Granularity	Frequency
Count	Cache Hit Count	`nv_cache_num_hits_per_model`	Number of response cache hits per model	Per model	Per request
	Cache Miss Count	`nv_cache_num_misses_per_model`	Number of response cache misses per model	Per model	Per request
Latency	Cache Hit Time	`nv_cache_hit_duration_per_model`	Cumulative time requests spend retrieving a cached response per model on cache hits (microseconds)	Per model	Per request
	Cache Miss Time	`nv_cache_miss_duration_per_model`	Cumulative time requests spend looking up and inserting responses into the cache on a cache miss (microseconds)	Per model	Per request

Similar to the Summaries section above for Inference Request Metrics, the per-model cache hit/miss latency metrics also support Summaries.

Note

For models with response caching enabled, the inference request summary metric is currently disabled. This is due to extra time spent internally on cache management that wouldn’t be reflected correctly in the end to end request time. Other summary metrics are unaffected.

Custom Metrics#

Triton exposes a C API to allow users and backends to register and collect custom metrics with the existing Triton metrics endpoint. The user takes the ownership of the custom metrics created through the APIs and must manage their lifetime following the API documentation.

The identity_backend demonstrates a practical example of adding a custom metric to a backend.

Further documentation can be found in the TRITONSERVER_MetricFamily* and TRITONSERVER_Metric* API annotations in tritonserver.h.

NVIDIA Triton Inference Server

Metrics

Contents